Neural network and regression approaches for predicting bubble point pressure in oil reservoirs

Abstract Petroleum engineers must accurately determine bubble point pressure (Pb) to optimize oil field development from discovery to abandonment. Errors in predicting this parameter can lead to inaccurate reservoir and production estimates. To enhance Pb predictions, A total of experimentally measured 1,161 oil samples from Egyptian oil reservoirs (604 heavy oil and 557 light oil) were used for regression analysis. In contrast, 232 datasets were used to support the artificial neural network model. Existing correlations for Egyptian crude oils of varying oil gravity demonstrated moderate accuracy, highlighting the need for improved estimation methods tailored to local reservoir conditions. This study revises existing literature correlations, which previously related Pb to reservoir temperature, flashed gas specific gravity, gas-oil ratio, and stock tank oil API gravity. In contrast, this work develops two new correlation models: one tailored for heavy oils (15.3 < API < 34.9) and another for light oils (35.6 < API < 53.1), providing improved accuracy for these distinct oil types. These models incorporate a compositional approach, emphasizing fluid composition and physical parameters for enhanced precision. The equations were revised using multiple linear regression analysis, resulting in a statistical coefficient of 99.08% for heavy oils and 96.49% for light oils. Additionally, Artificial Neural Network (ANN) models were trained on a comprehensive dataset to develop highly accurate mathematical models for predicting bubble point pressure.